Memory Graph Architecture (Phase 1)

Core Idea

Postgres + pgvector is the system of record for:

• Nodes (Event/Entity/Topic/Rule/Evidence)
• Edges (part_of/related_to/derived_from)
• Commit chain (append-only, hash-chained)

Graph traversal / spreading activation is done in the application layer.

Layering

• Raw / Evidence (immutable pointers):
  • stored as references in memory_nodes.raw_ref / memory_nodes.evidence_ref
• Normalized Nodes/Edges:
  • the stable substrate used for recall
• Derived artifacts:
  • Topics (from clustering), edge weights, summaries
  • should bump derivation_version so they can be recomputed

Rule Lifecycle

Rules are proposal-first:

• draft: created from extraction/abstraction; not used by default recall
• shadow: evaluated silently (produces suggestions but doesn't enforce)
• active: can influence planner/policy/tool selection
• disabled: kept for audit, excluded from recall

Per-run ephemeral override should live in execution state (not persisted).

Recall

1. Stage 1 (fast):
   • pgvector candidate retrieval from active tiers (hot + warm)
   • entity literal matches
   • rules in shadow/active
2. Stage 2 (smart):
   • fetch 1-2 hop neighborhood for seeds
   • spreading activation in app
   • output a bounded subgraph + citations (node/edge/commit/evidence refs)

Jobs

• topic clustering:
  • groups recent events into stable clusters
  • creates/updates Topic nodes + derived_from edges
• salience decay:
  • applies salience decay to active tiers
  • transitions by policy: hot -> warm -> cold -> archive
  • protects pinned/legal-hold nodes and enforces mutation budget